AI Is Drafting Disaster Alerts. Officials Fear the Public Won't Trust Them

A county emergency manager published an AI-generated hurricane graphic this month that showed a "sting jet" of high winds threatening the Alabama coast — and the map was not meteorologically accurate. Laura Myers, director of the University of Alabama's Center for Advanced Public Safety, recounted the incident after a meteorologist in Alabama called her following Hurricane Isaias, which made landfall on the Gulf Coast as a Category 3 hurricane. The emergency manager had used an AI chatbot to produce the graphic for his stakeholders.

"That county emergency manager was trying to get his stakeholders to focus on the risk in that county, and so he had asked AI to help him," Myers said. "It did what he wanted it to do, but it was not meteorologically accurate."

The episode illustrates the central tension now facing emergency management: AI can draft disaster alerts faster and often better than humans, but accuracy failures and eroding public trust carry life-and-death stakes. The stakes are rising as climate disasters grow more frequent and widespread.

How bad is human-written disaster communication?

Most emergency management agencies formally prohibit employees from using generative AI, yet Myers and other experts agree its use is likely already pervasive at state and local disaster departments. Hundreds of companies now market AI tools to emergency managers for tasks such as flood damage assessment and resource coordination, but public communication draws the most interest from government employees and startups.

The reason is simple: most emergency managers write poor alerts. They use jargon non-experts cannot parse, omit information on the assumption the public already knows it, and include irrelevant details that distract from what people actually need to know.

"A lot of agencies are really limited in resources and the number of people that they have," said Jeannette Sutton, a disaster researcher at the University of Albany who built a dashboard providing templates for disaster alerts. "I think they recognize it as a priority, but they don't prioritize it until they realize they should have. They're just trying to get their job done without screwing something up."

Emergency managers, not the Federal Emergency Management Agency, form the first line of defense after extreme weather at the state and local level. They often say "trust is our currency" — trust built and lost through public communication.

When does a bad alert kill people?

The record of flawed human communication is long:

• In 2016, adjacent counties in California issued a "voluntary evacuation" and a "mandatory evacuation" for the same wildfire.
• In Australia in 2019, officials warned of wildfires with the phrase "Watch and Act," which the government later said caused "significant confusion" because no one knew what it meant.
• In Montecito, California, weeks after the 2017 Thomas Fire, Santa Barbara County struggled to communicate mudslide risk to residents suffering what Sutton calls "warning fatigue." The county told everyone in the "voluntary evacuation zone" to "stay alert to changing conditions and be prepared to leave immediately." Most should have left at once. Many read the notice as optional; others refused to evacuate a second time in a month. The mudslide killed 23 people and injured more than 160.

"They're losing the opportunity to be that voice to help people to be safe," Sutton said of governments that issue bad warnings.

Why is alerting a natural fit for AI?

FEMA runs training sessions on alert best practices, but many officials forget their training when an emergency arrives. A large language model does not. It can read hundreds of thousands of alerts in minutes, digest the research on what works, and edit a human alert in seconds.

"If you've only sent one [alert] in your life, and that was 15 years ago, and you need to remember all your training, and you're thinking 'Hey, there's a chemical plume coming my way, the phone's ringing, the radio's blaring, I need help,' the AI is probably gonna do a better job than you are," said Rob Dale, a former meteorologist who now runs the emergency management department in Ingham County, Michigan.

Dale has used AI to draft practice alerts behind the scenes but has never sent one. Ingham County, a small central Michigan county, faces no hurricanes or major fires. If such an event struck, Dale said he would let a chatbot draft his incident communication but keep control over sending the alert.

How good is the EM1 model in practice?

EM1, a startup producing AI software for emergency managers, trained a model on Sutton's research. Its executives say the model already beats almost any human emergency manager. The company has tested it in-house on scenarios including wildfires and tsunamis, and it matches or exceeds human capabilities on most benchmark tests.

Sutton demonstrated the software on a real alert from a New Mexico county warning residents about flooding on the San Juan River. The original read: "Breeched banks. The water is rising. EVACUATE NOW. McGee Park open for people and pets, RV."

The EM1 model produced a revision in seconds: "San Juan County: Floodwater is rising near [location]. People in [evacuation area]: LEAVE NOW for higher ground. Do not walk or drive through floodwater. McGee Park is open for people and pets. Check [official update source] for updates."

The model identified and fixed several problems. The original text does not say who is sending the alert, name the specific flooding location, or clarify who must evacuate and how to do so safely. It misspells "breached" and offers no guidance on where to find further information.

Are agencies actually deploying it?

EM1 holds contracts with around a hundred government agencies and school districts. None uses the software to automate extreme weather alerts yet; they use it only to workshop plans and internal preparations.

Yet every expert interviewed for this story said many emergency managers already use chatbots to write messages, even if they will not admit it. Sutton said managers in her alert training sessions often present high-quality alerts drafted in seconds with ChatGPT or Gemini.

"Official adoption is pretty low, but people are unofficially using it, absolutely," said Sarah Miller, a professor of emergency management at Pierce College in Florida and a longtime instructor at the National Disaster Preparedness Training Center. "They're using their own personal accounts, or they're using free ones."

What could go wrong?

Even researchers bullish on chatbots warn the switch could backfire. Members of the public skeptical of AI or distrustful of government may spurn agencies that publish AI-generated graphics or social posts. In Syracuse, New York, a union of public employees sued the local government for using an AI service to answer non-emergency 911 calls.

"We trust things that we perceive [are] authentic and care about us," said Ginny Katz, founder of HazAdapt, a consultancy that helps disaster officials deploy technology. "AI-generated content does not feel authentic."

Chatbots could also let agencies send more alerts, faster — increasing the risk of warning fatigue, the same phenomenon that contributed to the Montecito deaths.

As elsewhere in AI deployment, experts say the software is only as capable as its user. A seasoned manager can accelerate communication with chatbot-drafted alerts; an inexperienced one rushing an urgent alert cannot evaluate the output of a model like ChatGPT or Claude. That limitation has not stopped understaffed departments from using the technology, and experts say it is too early to tell how AI-generated content is affecting public trust in government.

"If you had a human sending you a text message versus a computer sending you a text message, which one would you believe is being more accurate?" said Justin Snair, founder of Preppr, another company selling AI software to emergency managers.

The answer, he said, depends on the person receiving it — which means every alert, human or machine, now carries an added test.